Turn Long Text into Images to Save Tokens? Snapcompact Halves LLM Costs

arpit_bhayani · x · 2026-08-03

Stencil introduced Snapcompact, a novel context compaction technique: instead of summarizing text when the context window fills up, it renders the text into dense pixel-font bitmaps (PNG images) and feeds them back to the model.

Core Mechanism & Data:

A 1568×1568 PNG can hold about 40,000 characters (10,000 tokens).
Multimodal models read these images at roughly one-third the input price of text.
Tests on models like Claude, GPT-5.5, and Gemini show that this method reduces total inference costs by about 50% or more at F1 parity, while avoiding the information loss or model abstentions common with traditional text summarization.

Original post →

Turn Long Text into Images to Save Tokens? Snapcompact Halves LLM Costs

More from Infra